Agent

An agent is a software system that controls how an inference request is composed during loop execution.

It prepares the inference requests sent to the model, combining instructions, available tool definitions, and context from the loop. The model's output then informs the next step: execute tools, request another inference, or finish the task.

The agent can also request a human decision through human-in-the-loop or transfer responsibility to another agent through a handoff.

An agent can maintain memory across tasks. The context belongs to the loop, while relevant information from the agent's memory can be added to that context for use during inference.